Deep Reinforcement Learning for Analog Circuit Sizing
Zhenxin Zhao, Lihong Zhang
Abstract
Automated analog circuit sizing is always a challenging task, due to high complexity involved, huge design space searched, and conflicting constraints traded off. This paper proposes an automated trial and error approach that combines reinforcement learning with deep learning for analog circuit sizing. Through the self-improvement learning way, the proposed method behaves like a designer, who learns from trials and derives experience, evolving itself to finally discover the sizes that satisfy the performance specification based on simulation results. In order to greatly reduce the number of simulations, we propose a symbolic filter that builds a polynomial equation system by utilizing the curve-fitting results and then applies the worked out small-signal parameter values to implement symbolic analysis to quickly evaluate the circuit performance, passing only the satisfied ones to the simulator. Our experimental results demonstrate the reliability of the proposed method, and also reveal the self-improvement capability.